What if the phrase used to shut a conversation down is only ever reached for by the side that’s losing it?
“Correlation doesn’t equal causation” sounds like a rule of science. It isn’t. It’s a reminder — one line from a much longer conversation about how scientists actually decide whether a pattern is meaningful. Strip it out of that context and hand it to someone as a conversation-ending card, and you’ve turned a caveat into a weapon.
So here’s the question worth sitting with: when does that card get played, and when does it stay in the deck?
The Bradford Hill Viewpoints: The Conversation Everyone Skips
Epidemiologists have had a real answer to “is this correlation meaningful?” since 1965. Sir Austin Bradford Hill — the same statistician whose work helped establish that smoking causes lung cancer — proposed nine viewpoints for weighing whether an association is likely causal: the strength of the association, whether it’s consistent across different populations and studies, whether exposure clearly precedes the outcome, whether there’s a dose-response relationship, whether there’s a plausible biological mechanism, and more.
Bradford Hill’s own point was that no single study proves causation, and that correlation without further work shouldn’t be mistaken for it. But the tool exists precisely so that people can go further than “correlation isn’t causation” — so that a pattern can be tested, weighed, and either strengthened or discarded on its merits. That’s the part that tends to go missing whenever the phrase gets deployed as a full stop.
How the Tobacco Industry Used “Correlation Isn’t Causation” to Delay Action on Smoking
This isn’t a new trick. In the 1950s, as case-control studies from Doll and Hill, and separately from Wynder and Graham, showed a clear association between smoking and lung cancer, many influential voices argued that the evidence showed “only correlation” — not causation. Internal industry documents later showed some tobacco company scientists privately agreed smoking caused cancer years before their companies’ public position changed. The 1954 “Frank Statement to Cigarette Smokers,” issued by tobacco companies, publicly insisted there was “no proof that cigarette smoking is one of the causes” of cancer — a position maintained long after the internal science had moved on.
It took until 1957 for a joint scientific panel to state plainly that the evidence established beyond a reasonable doubt that smoking was a cause of lung cancer, and until the 1960s for governments to say so formally. In the gap between “here’s a strong, consistent, dose-responsive correlation” and “we accept this is causal,” the phrase doing the heavy lifting for the other side wasn’t science. It was delay.
Where Correlation Is Treated as Enough: Climate Science and Drug Safety
Now hold that up against fields where the phrase rarely ends the conversation.
Climate science runs substantially on correlational and observational data — ice cores, temperature records, ocean measurements — layered with modelling. Mainstream scientific assessment does not require a randomised controlled trial with a control-group Earth before accepting the evidence as sufficient for action.
Pharmacovigilance — the entire discipline of monitoring drugs after they’re approved — is built on detecting patterns between medicines and adverse events in the real world, then investigating those signals and, where appropriate, acting on them long before every biological mechanism is fully understood.
Public health policy routinely moves on correlational evidence: population-level associations between a risk factor and an outcome are treated as a reasonable basis for guidance long before a lab has mapped the exact biological pathway.
In each of these areas, correlation is treated as informative — not proof on its own, but a legitimate signal worth acting on, refining, and building policy around. The four-word dismissal is far less common.
Glyphosate and Cancer: Where the Phrase Gets Used Instead of the Evidence
Compare that with glyphosate.
In 2015, the International Agency for Research on Cancer (IARC) — the WHO’s cancer research arm — classified glyphosate as “probably carcinogenic to humans” (Group 2A), based on limited human epidemiological evidence, mainly for non-Hodgkin lymphoma, combined with what it judged to be sufficient evidence of carcinogenicity in animal studies and strong evidence of genotoxicity. Several later meta-analyses of case-control studies out of the US and Europe have consistently found a moderate increase in non-Hodgkin lymphoma risk associated with glyphosate exposure, particularly among people exposed on multiple days a year.
Regulators in most other jurisdictions — the US EPA, EFSA, Health Canada, Australia’s APVMA, and New Zealand’s own EPA among them — reached the opposite conclusion, determining that the available evidence did not support classifying glyphosate as carcinogenic to humans under their respective assessment frameworks. It’s worth being straight about why: these agencies weighed a larger set of industry-submitted toxicology studies, not all of which are publicly available, and applied different thresholds for what counts as sufficient genotoxicity evidence. One peer-reviewed comparison of the underlying genotoxicity assays found that a majority of the published, publicly available studies reported a positive genotoxic signal, while the registrant-submitted studies EPA leaned on did not — which is precisely the kind of divergence Bradford Hill’s framework exists to interrogate, rather than wave away.
This is a genuine, live scientific dispute between serious institutions using different evidence bases and different rules for what counts as “enough.” That’s a fair thing to say plainly. What’s not fair is treating “correlation isn’t causation” as though it settles that dispute in one direction — when many of those same regulators have, in other contexts, built entire frameworks on correlational signals they judged strong enough to act on.
The Real Standard: One Rule, Applied Every Time
That’s the pattern worth naming: the phrase isn’t applied consistently as a scientific standard. It’s applied selectively, as a rhetorical circuit-breaker — deployed hardest against the correlations that would require inconvenient action, and quietly absent from the correlations everyone’s already comfortable acting on.
A generation has been handed a four-word stopper and told it was rigor. Actual rigor looks like Bradford Hill’s nine viewpoints — strength, consistency, temporality, dose-response, plausibility, and the rest — applied evenly, to every claim, regardless of who’s inconvenienced by the answer.
So next time someone reaches for the phrase, the question worth asking back isn’t “is that true?” It obviously is — correlation genuinely isn’t causation. The question is: are you applying that standard everywhere, or only here?
Note: this piece describes a genuine, ongoing scientific and regulatory disagreement about glyphosate’s carcinogenicity. Readers should draw their own conclusions from the primary sources linked throughout.
Further Reading
What if the four-word dismissal isn’t really about statistics at all — but about who gets to decide when a pattern is finally taken seriously? Every source below was written by someone who went further than the phrase allows: historians who traced the doubt back to its source, a journalist who followed the paper trail, scientists who ran the numbers, and a corporation that has now agreed to pay billions of dollars to resolve litigation over a substance it has consistently maintained does not cause cancer.
Merchants of Doubt*
Naomi Oreskes and Erik Conway’s history of how the same small network of scientists and PR strategists ran the doubt campaign for tobacco, then reused it for acid rain, the ozone hole, and climate change. The direct throughline from this article’s tobacco case study to today’s playbook.
*For your convenience, we provide links to Amazon.com. If you choose to purchase through these links, we may receive a small commission — at no additional cost to you. Your support helps us continue our work.
Whitewash: The Story of a Weed Killer, Cancer, and the Corruption of Science [A review]
Investigative journalist Carey Gillam’s book-length account of Monsanto’s decades-long effort to shape the science and regulatory record on glyphosate, drawn from internal industry documents and interviews with affected farming families.
The Truth (and Lies) of Correlation vs. Causation
A Scientific American explainer on the formal logical fallacy behind the phrase, with everyday examples of when correlation misleads — and a reminder that correlation is sometimes genuinely all researchers have to work with.
“Improving the Teaching of ‘Correlation Does Not Equal Causation’ in Introductory Psychology”
A 2025 peer-reviewed paper arguing the phrase is widely misunderstood by students to mean correlation can never indicate causation, and proposing better ways to teach the actual, more nuanced rule.
Zhang et al., “Exposure to Glyphosate-Based Herbicides and Risk for Non-Hodgkin Lymphoma: A Meta-Analysis and Supporting Evidence” (2019)
The peer-reviewed meta-analysis that found a 41% increased relative risk of non-Hodgkin lymphoma among the most highly glyphosate-exposed individuals, combining case-control and cohort data including the US Agricultural Health Study.
Bayer’s announcement of its $7.25 billion Roundup class settlement (March 2026)
Bayer’s own statement on the nationwide settlement covering current and future non-Hodgkin lymphoma claims linked to Roundup, alongside billions more in separate resolutions — a live illustration of how unresolved the “settled science” claim really is.
None of these sources agree on everything. That’s the point. Read them side by side and ask the same question this article keeps returning to: is the disagreement really about the evidence — or about who’s inconvenienced by what it shows?
Image Source & Attribution
The feature image on this page was created using AI-assisted image generation from an original concept developed by No More Glyphosate NZ and refined for publication in Canva.
AI is a useful creative tool for visualising complex investigative topics that cannot be meaningfully photographed, allowing us to illustrate ideas without implying that any specific scene or event actually occurred.
